Signal Processing - November 2016 - 51

safety risks. Research shows that glancing away from the road
applications, languages, or modalities. It also intends to show
for two seconds or longer may increase the risk of an accident
the capability of learning dialog strategies to improve comfrom four to 24 times [12]. Such research evidence has led some
munication with drivers when an ASR engine make errors. A
original equipment manufacturers to prohibit the use of some of
statistical framework of a partially observable Markov deciinterfaces with hierarchical menu structures during driving. The
sion process was used to address uncertainty in speech recogvoice-based systems have been shown to reduce driver look-away
nition results [11]. Despite different foci of these research
time and lessen spatial and cognitive demands on the driver. It
projects, they share many important building blocks in an
is expected that recent advancement in speech recognition techadvanced speech dialog system, as shown in Figure 1.
nology; other sensing technologies such as gesture on touch pad,
Influential automotive infotainment HMI products on
on steering wheel surface, and in air; as
the market include BMW iDrive, Mercedes
well as deeper modality fusion technoloCOMAND, Audi MMI, Ford Sync, GM CUE,
In the past two decades,
gies would greatly liberate the designers so
Lexus Enform, and Acura AcuraLink. In genervehicles equipped with
that many new devices and services can be
al, these HMI products are used to control mulincorporated without many physical contiple in-car functionalities including navigation,
both human and vehicle
straints in the cockpit [13], [14].
entertainment, multimedia, telephony, vehicle
sensors have been
dynamics, and so on. Existing HMI input techdeployed to collect
nologies fall into two categories: haptic based
Related intelligent assistant
realistic data on drivers,
and voice based. Of haptic-based input methtechnologies
vehicles, and driving
ods, the two most common are control knobs
Likewise, many efforts were devoted to
environments for the
and touch screens. German cars, e.g., BMW
the field of general-purpose voice-enabled
iDrive and Audi MMI, typically use control
intelligent personal assistants (IPAs). The
development of humanknobs. While these control knobs are simple
U.S. Defense Advanced Research Projcentered technologies.
to operate, it often takes multiple steps to
ects Agency (DARPA) funded a number
access a function deep within a menu. Thus,
of key IPA-related programs. The
these interfaces impose cognitive demands on drivers, requiring
DARPA Communicator project was a major early effort in
them to memorize where menu items are located and navigate
developing robust multimodal speech-enabled dialog systems
through the menu while driving. In contrast, Japanese cars tend
with advanced conversational capabilities for engaging human
to use touch screens. Navigating touch-screen interfaces can be
users in mixed-initiative interactions. From 2003 to 2008,
more intuitive and efficient, but they impose a spatial limitaDARPA funded the Cognitive Agent that Learns and Organiztion on drivers, as they must reach out to different areas of touch
es (CALO) project. CALO attempted to build an integrated sysscreen to access predesignated buttons in a shallower but still
tem capable of true artificial intelligence's (AI's) key features
hierarchical menu structure. As both knobs and touch screens
such as the ability to learn and adapt in adverse situations, and
require hand-eye coordination, haptic interfaces pose inherent
comfortably interact with humans [15]. The CALO project had

Knowledge
Acquisition

Knowledge
Base

Knowledge
Management

Application Management
Phone

Speech
Recognition

MP3
Player

Natural
Language
Understanding
Dialog
Management

Speech
Synthesis

Navigation

Database

Response
Generation

FIGURE 1. A generic block diagram for a typical in-vehicle spoken dialog system.
IEEE SIgnal ProcESSIng MagazInE

November 2016

Table of Contents for the Digital Edition of Signal Processing - November 2016